A single inference run is not one model call. It is a four-stage pipeline across two specialized model heads. The vision head (27B) reads the image and produces structured findings. The text head (20B) turns those findings into an impression, differential, and management plan. Together they compose OmniBioFex 1.0 47B.
Every packet you see is one token streaming along the rail. Hover any chamber for its role.
Click any stage. Watch how a single image travels through the pipeline and becomes a structured report.
A real-time peek at what a running inference actually looks like on the backend.
Vision and reasoning are different problems. A single monolithic model is a compromise on both. So we split them.
Extracts 6–12 discrete findings with per-item confidence and 3×3 region tags. Runs a parallel fracture-hunt audit. Emits strict JSON.
Takes the structured findings, writes the one-line impression, ranks a differential diagnosis, and drafts 3–6 management steps.
Each stage calls the head that is best for it. Routing beats stacking — predictable memory, predictable cost.
Four entry points to the same two-headed pipeline. Each one chooses which heads to call, and how many times.
Vision head → text head (impression + differential). One pass each. Optional 2-pass region verification.
→ ONE VISION PASS · ONE REASON PASSTwo vision passes with agreement scoring. Text head reconciles the two into a single report.
→ TWO VISION PASSES · AGREEMENT SCOREVision head in bounding-box mode. Returns pixel-space rectangles for every candidate lesion.
→ VISION ONLY · BOUNDING BOXESTwo vision passes (prior + current) then the text head produces an interval-change report.
→ PRIOR + CURRENT · INTERVAL REPORTThe engine runs live in the chat — attach a scan and watch the report assemble.